Research On Pedestrian Detection And Re-identification Based On Deep Learning

With the accelerated construction of smart cities,deep learning technologies have significantly played their advantages in related fields.Facing the growing demand of public safety,pedestrian detection and re-identification plays a crucial role as a humancentered computer vision task.Currently,pedestrian detection and re-identification technologies are being widely used in public security criminal investigation,humancomputer interaction,intelligent transportation and other fields.In modern cities,large-scale surveillance equipment provides a large amount of surveillance video data,which provides a reliable basis for tracking specific target persons.It also places higher demands on the performance and accuracy of pedestrian detection and re-identification technologies,and requires continuous technological innovation and optimization to meet these challenges.This paper aims to study the problems encountered in pedestrian detection and re-identification under real scenarios,such as: in the pedestrian detection process,pedestrian features are affected by different scales,background confusion,occlusion and other factors,resulting in insufficient feature characterization ability,which affects the detection effect;The existing two-stage anchor-based network detection model has high accuracy but too slow,and the use of single-stage network will lead to a large number of negative In the process of pedestrian re-identification,there are cross-device and multi-angle pedestrian features,and the model that extracts global features or local features singularly cannot achieve the ideal recognition effect.The research work of this paper includes:1.Improving the single-stage full convolutional neural network and proposing a feature-enhanced pedestrian detection method based on anchor-free neural network,combining the features extracted from different depths and designing a featureenhanced module to improve the pedestrian feature characterization capability.A key central region judgment mechanism is added to the feature extraction network to reduce the generation of negative samples,reduce the negative effect on the model,and improve the model detection performance.2.Design a multi-branch joint learning pedestrian re-identification model.Different branches have different division of labor to extract global features and local features of pedestrians respectively.Considering the different characteristics displayed by the two kinds of features and the phenomenon that unaligned pedestrian features will lead to performance degradation,the feature alignment module is designed to transform different types of feature maps to achieve spatial alignment,enhance the model performance and improve the recognition accuracy of the model.The pedestrian detection and re-identification models are pre-trained using large datasets publicly available in the domain,respectively,and the backbone networks are pre-trained using Image Net.The experimental results are analyzed and compared with the rest of the models,showing that both models have performance improvements.The pedestrian detection model,with an accuracy result of 94.16%,improves the detection accuracy by 6.8% compared to the benchmark model;The pedestrian re-identification model achieves 87.64% and 77.62% accuracy results on two large datasets,respectively,and the experimental results are competitive compared to other re-identification models.